Correlating neuronal activity and large volume nanoscale imaging using AI

Understanding how the brain works - a major driving force for the development of AI - requires knowledge of the wiring of its individual neural circuits. This can be achieved using electron microscopy to image large volumes of brain tissue and then map the connections between neurons (synapses) at the nanometre-scale. However, while the 'wiring diagram' that results from this effort is necessary to understand the brain, it is not, by itself, sufficient. We also need information about the function of each connection - i.e., how effective each synapse is at information signalling.

In this proposal, our key aim is to complement brain circuit wiring diagrams ('connectomes') with a readout of synaptic activity, allowing us to better understand the relationship between brain structure and function in health and disease.

We recently developed an experimental technique that directly addresses this challenge, allowing measurements of strength and activity of all synapses within a brain volume. This is achieved using a special type of electron microscopy that yields functional brain maps with extraordinary resolution. However, these datasets are very large and almost impossible for humans to analyse at scale. To solve this, in collaboration with an expert partner in AI for image analysis, Jan Funke (Janelia Research Campus-USA), we have developed powerful machine-learning approaches that can automate the readout of wiring and functional properties of these circuits.

Our experimental approach involves introducing a special marker into synapses of target brain circuits to read out functional information. In order to describe structure-function relationships in existing datasets from neural circuits that do not include this marker, we will build on the close links between AI and connectomics -fully aligned with the BBSRC remit and the scopes of this call - to create a new AI-based tool that will enable us to assess synaptic function in connectomes based on structure alone. This means that the many wiring diagrams already collected by researchers across the world could be re-analysed to add crucial functional information. To achieve this exciting objective, our labelled experimental data will act as the 'ground truth', and AI networks will be trained on this dataset to learn how to relate structural characteristics of synapses to functional properties.

This approach will provide a major advance in the field: the larger the tissue volume, the more compelling is the need to develop and optimise new automated tools to accelerate discovery. Our approach will thus be transformative for the many researchers interested in generating functional maps of circuits in the brain. By sharing data and methodology, we will contribute to the field of connectomics and make it more equitable and accessible to the broader scientific community. This collaborative culture will reach beneficiaries that would not otherwise be able to capitalise on such data and methodology because of economical disadvantages and/or lack of access to the technology requiredfor such experiments. It will benefit both the neuroscience and AI communities, and will add enormous value to the vast existing datasets available globally, thus deepening our understanding of how biological and artificial neural networks operate.

Ultimately, our newly developed AI tools will also have the potential to predict function from structure in medical images, which could support and facilitate diagnoses, improve outcomes, widen the impact of this partnership's work to translational fields and make a positive-impact on the community's welfare.